(perf): GPU parity — depth-tagged touched-others stack, PoolCheckpointState, lookup memo (CUDA/Metal)

[mgyoo86]

Problem

The GPU extensions still ran the pre-#50 design: _lazy_checkpoint! eagerly snapshotted every registered fallback type at scope entry (O(registered types) — the pollution bug #50 fixed on CPU), lazy rewinds iterated the whole values(pool.others) dict, and every fallback acquire paid an IdDict lookup. #51's reorder deferred GPU parity so both backends could mirror the final design once.

Fix

Mirror #51's architecture in both extensions, sharing one implementation where possible:

1. Shared stack helpers — _touch_fallback_pool! / _drain_touched_others! / _truncate_touched_others! generalized over AbstractArrayPool (+ _check_level); CPU, Metal, and CUDA now run the same code. Compile-time _runtime_check gating keeps the S=0 DCE contract (verified via @code_typed: zero dynamic dispatch, branch fully folded).
2. PoolCheckpointState on GPU typed pools — MetalTypedPool/CuTypedPool nest their checkpoint fields into the same concrete struct as CPU (property forwarding keeps all call sites source-compatible), so the S=0 drain runs zero-dispatch on Vector{PoolCheckpointState}.
3. Depth-tagged stack + lookup memo — eager snapshots deleted; producers push one (state, depth) entry per first touch; lazy rewinds drain by depth tag; full checkpoint/rewind keep their eager-sweep + truncate pairing. get_typed_pool! memoizes the last fallback lookup.

Float16 caveat (GPU-only divergence): Float16 is a fixed slot with _fixed_slot_bit == 0 (bit-7 reassignment), so stack routing uses !(T <: _*_FIXED_TYPES) — never the bit test. Routing it onto the stack would double-rewind against the lazy rewinds' Float16 special case and corrupt parent-scope counts (caught in plan review; regression testsets guard it on both backends).

Also: dead _make_cuda_pool(level, old)/_transfer_cuda_pool removed (zero callers; same rationale as #51's removal — and its positional constructor would no longer compile). Surfaced along the way and fixed: GPU poison-fill now mirrors the CPU best-effort contract (never throws during rewind on zero-less isbits types), and GPU/auto-manage tests are runtime_check=1-clean (S-adaptive asserts; scopes never return pool-backed arrays).

Benchmarks

Metal (Apple Silicon, local), lazy @with_pool :metal scope acquiring one fallback type — same method as #50/#51:

	1 registered type	8-type/1-type ratio
master	~390 ns	8.2× (eager snapshot is O(others))
this PR	~105 ns	≈1.00 flat for others = 1..13

Same-process registry-size-isolated sweep is the signal (a fresh-process JIT ordering artifact adds a ~50 ns constant offset unrelated to registry size — details in the branch notes). Zero-dispatch drain confirmed on the S=0 path.

Testing

• test/metal/test_touched_others.jl + CUDA mirror (~850 lines): pollution regression, f_outer exception-leak recovery, depth-tag exactness/monotonicity, memo lifecycle (empty! clears / reset! preserves), Float16 non-interaction (lazy, typed, mixed), R=1 three-stack lockstep (happy path, exception unwind, nested depths), mixed fixed+fallback tuples.
• Metal: full suite green on-device. CPU: full suite green (shared-helper refactor pinned by the existing touched-others + zero-alloc testsets).
• CUDA: line-audited transcription of the reviewed Metal diff (normalized cross-diff, syntax-gated) — functional run on CUDA hardware done separately before merge.

Follow-ups (tracked)

• PR4 macro work: Foo{Float64} static widening, per-scope tp hoisting, and an expansion-time escape lint for the two syntactic footguns (direct acquire / x .= v tails).

[codecov]

Codecov Report

❌ Patch coverage is 57.14286% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.81%. Comparing base (bc657a1) to head (760d0fd).

Files with missing lines	Patch %	Lines
src/state.jl	50.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #52      +/-   ##
==========================================
- Coverage   95.83%   95.81%   -0.03%     
==========================================
  Files          16       16              
  Lines        3677     3678       +1     
==========================================
  Hits         3524     3524              
- Misses        153      154       +1     

Files with missing lines	Coverage Δ
src/types.jl	87.23% <100.00%> (+0.13%)	⬆️
src/state.jl	96.64% <50.00%> (-0.20%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.